Manifesto
The inference behind every AI workload makes a trade-off between latency and throughput. In the first iteration of generative AI, systems optimized for low latency return tokens and output as quickly as possible to a user waiting on the other end.
But speed comes at a cost. A large and growing share of AI work isn’t waiting on a human at all. Asynchronous use cases—like deep research, code review, security review, evals, and embeddings—require agentic pipelines that spend hours running in the background, without humans in the loop. In this paradigm, shaving milliseconds off a single response buys you nothing.
Optimizing for latency leaves compute underutilized. Optimizing for throughput requires a fundamental rethinking of how to consume resources most efficiently, driving utilization up and cost per token down. From the lowest levels of the stack starting from silicon, all the way up to sandboxes, Sail’s platform is designed as one system purpose-built for long-running tasks, so agents persist for hours and days rather than dying between calls.
Using Sail’s inference API, token budgets can now go 10x further compared to other providers. Sail offers a model-agnostic, open endpoint with elastic provisioning. Our API can spin up and down workloads from 0 to trillions of tokens in a matter of minutes, with a robust control plane delivering reliable service over unreliable compute. Developers no longer need to worry about rate limits, and can run workloads scalably, reliably, and consistently.
Sailboxes are the most efficient cloud environment for agents, making workflows over 3x cheaper. Our novel architecture ensures users only pay for the portion of CPU, memory, and disk their agent actually uses, with automatic sleep during inference. Our customers can build all their agents in a Sailbox and have them live forever on the cloud, without worrying about cost or reliability.
With Sail, your agents can run trajectories with more turns and richer context. You can serve more users with the same margin, and have room to experiment without rationing tokens. Your workloads can withstand retries, failures, and any other error correction at the hardware level within Sail’s infrastructure.
Sail’s platform pushes the frontier of what can be possible in an agent abundant world by maximizing your intelligence per dollar.
Join us
We are systems nerds with commercial focus, who’ve worked at companies like Together AI, Apple, Yubico, Jane Street, Robinhood, and more. We:
- Write kernels to push towards speed-of-light performance on GPUs
- Use unusual parallelism schemes and scheduling techniques, which have never been run at scale before, in our inference engine
- Distribute work globally to workers to maximize robustness and fleet utilization, while tolerating arbitrary and immediate failure of any worker
- Run a production service using a small but mighty team of owners
We all love the craft of engineering and the pursuit of peak performance. If that sounds like you, then join us.
